Google Workloads for Consumer Devices: Mitigating Data Movement Bottlenecks Amirali Boroumand

Transcription

1 Google Workloads for Consumer Devices: Mitigating Data Movement Bottlenecks Amirali Boroumand Saugata Ghose, Youngsok Kim, Rachata Ausavarungnirun, Eric Shiu, Rahul Thakur, Daehyun Kim, Aki Kuusela, Allan Knies, Parthasarathy Ranganathan, Onur Mutlu

2 Consumer Devices Consumer devices are everywhere! Energy consumption is a first-class concern in consumer devices 2

3 Popular Google Consumer Workloads Chrome Google s web browser TensorFlow Mobile Google s machine learning framework Video Playback Google s video codec Video Capture Google s video codec 3

4 Energy Cost of Data Movement 1 st key observation: 62.7% of the total system energy is spent on data movement SoC Data Movement CPU CPU CPU CPU L1 L2 Compute Unit DRAM Processing-in-Memory (PIM) Potential solution: move computation close to data Challenge: limited area and energy budget 4

5 Using PIM to Reduce Data Movement 2 nd key observation: a significant fraction of data movement often comes from simple functions We can design lightweight logic to implement these simple functions in memory Small embedded low-power core PIM Core Small fixed-function accelerators PIM Accelerator PIM PIM Accelerator Accelerator Offloading to PIM logic reduces energy by 55.4% and improves performance by 54.2% on average 5

6 6 Goals 1 2 Understand the data movement related bottlenecks in modern consumer workloads Analyze opportunities to mitigate data movement by using processing-in-memory (PIM) 3 Design PIM logic that can maximize energy efficiency given the limited area and energy budget in consumer devices

7 7 Outline Introduction Background Analysis Methodology Workload Analysis Evaluation Conclusion

8 8 Potential Solution to Address Data Movement Processing-in-Memory (PIM) A potential solution to reduce data movement Idea: move computation close to data Reduces data movement Exploits large in-memory bandwidth Exploits shorter access latency to memory Enabled by recent advances in 3D-stacked memory

9 9 Outline Introduction Background Analysis Methodology Workload Analysis Evaluation Conclusion

10 10 Workload Analysis Methodology Workload Characterization Chromebook with an Intel Celeron SoC and 2GB of DRAM Extensively use performance counters within SoC Energy Model Sum of the energy consumption within the CPU, all caches, off-chip interconnects, and DRAM CPU CPU L1 L2 DRAM

11 11 PIM Logic Implementation SoC DRAM Logic Layer PIM Core Customized embedded general-purpose core No aggressive ILP techniques 256-bit SIMD unit PIM Accelerator PIM PIM Accelerator Accelerator N Small fixed-function accelerators Multiple copies of customized in-memory logic unit

12 Workload Analysis Chrome Google s web browser TensorFlow Google s machine learning framework Video Playback Google s video codec Video Capture Google s video Codec 12

13 Workload Analysis Chrome Google s web browser TensorFlow Google s machine learning framework Video Playback Google s video codec Video Capture Google s video codec 13

14 14 How Chrome Renders a Web Page Loading and Parsing Painting HTML CSS HTML Parser CSS Parser Render Tree Layout Rasterization Composit- Compositing

15 15 How Chrome Renders a Web Page Loading and Parsing Layouting Painting HTML CSS HTML Parser CSS Parser Render Tree Layout assembles all layers into a final screen image Rasterization Compositing calculates the visual elements and position of each object paints those objects and generates the bitmaps

16 16 Browser Analysis To satisfy user experience, the browser must provide: Fast loading of webpages Smooth scrolling of webpages Quick switching between browser tabs We focus on two important user interactions: 1) Page Scrolling 2) Tab Switching Both include page loading

17 Scrolling 17

18 What Does Happen During Scrolling? HTML CSS rasterization uses color blitters to convert the basic primitives into bitmaps HTML Parser CSS Parser Render Tree Layout Color Blitting Texture Tiling Rasterization itng Compositing to minimize cache misses during compositing, the graphics driver reorganizes the bitmaps 18

19 19 Scrolling Energy Analysis Frac%on of Total Energy 100% 80% 60% 40% 20% 0% Google Docs Texture Tiling Color Bli6ng Other Gmail Google Calendar Word- Press Twitter Animation AVG 41.9% of page scrolling energy is spent on texture tiling and color blitting

20 Total Energy (pj) Scrolling a Google Docs Web Page Texture Tiling Color Bli6ng Other CPU L1 LLC Interconnect Mem Ctrl A significant portion of total data movement comes from texture tiling and color blitting DRAM Frac%on of Total Energy 37.7% of total system energy 40% 30% 20% 10% 0% 77% of total energy consumption goes to data movement Data Movement Texture Tiling Compute Color Bli6ng 20

21 Total Energy (pj) Scrolling a Google Docs Web Page Texture Tiling Color Bli6ng Other A significant portion of total data movement comes from texture tiling and color blitting 37.7% of total system energy 40% 30% 20% 10% 0% 77% of total energy consumption goes to data movement CPU L1 LLC Interconnect Mem DRAM Ctrl for texture tiling and color Data Movement blitting? Compute Can we use PIM to mitigate the data movement cost Frac%on of Total Energy Texture Tiling Color Bli6ng 21

22 Can We Use PIM for Texture Tiling? CPU-Only!me CPU Memory CPU + PIM CPU PIM Texture Tiling Rasterization Read Bitmap Linear Bitmap Rasterization idle Conversion high data movement Invoke Compositing Write Back Texture Invoke Tiles Compositing Major sources of data movement: 1) poor data locality 2) large rasterized bitmap PIM execution size (e.g. 4MB) Texture tiling is a good candidate for Linear Bitmap Conversion Texture Tiles 22

23 Can We Implement Texture Tiling in PIM Logic? Linear Bitmap Texture Tiling Tiled Texture Requires simple primitives: memcopy, bitwise operations, and simple arithmetic operations PIM Core PIM Accelerator 9.4% of the area available for PIM logic 7.1% of the area available for PIM logic PIM core and PIM accelerator are feasible to implement in-memory Texture Tiling 23

24 24 Color Blitting Analysis Generates a large amount of data movement Accounts for 19.1% of the total system energy during scrolling Color blitting is a good candidate for PIM execution Requires low-cost operations: Memset, simple arithmetic, and shift operations It is feasible to implement color blitting in PIM core and PIM accelerator

25 25 Scrolling Wrap Up Texture tiling and color blitting account for a significant portion (41.9%) of energy consumption 37.7% of total system energy goes to data movement generated by these functions 1 Both functions can benefit significantly from PIM execution 2 Both functions are feasible to implement as PIM logic

26 Tab Switching 26

27 27 What Happens During Tab Switching? Chrome employs a multi-process architecture Each tab is a separate process Chrome Process Tab 1 Process Tab 2 Process Tab N Process Main operations during tab switching: Context switch Load the new page

28 28 Memory Consumption Primary concerns during tab switching: How fast a new tab loads and becomes interactive Memory consumption Chrome uses compression to reduce each tab s memory footprint CPU Compressed Inactive Tab Tab Decompression Compression DRAM ZRAM

29 29 Data Movement Study To study data movement during tab switching, we emulate a user switching through 50 tabs We make two key observations: 1 2 Compression and decompression contribute to18.1% of the total system energy 19.6 GB of data moves between CPU and ZRAM

30 30 Can We Use PIM to Mitigate the Cost? compression CPU CPU-Only Memory Swap out N pages Read N Pages Compress Write back Other tasks Uncompressed Pages high data movement ZRAM!me CPU CPU + PIM PIM Swap out N pages Other tasks No off-chip data movement Uncompressed Pages Compress ZRAM PIM core and PIM accelerator are feasible to implement in-memory compression/decompression

31 31 Tab Switching Wrap Up A large amount of data movement happens during tab switching as Chrome attempts to compress and decompress tabs Both functions can benefit from PIM execution and can be implemented as PIM logic 2

32 Workload Analysis Chrome Google s web browser TensorFlow Google s machine learning framework Video Playback Google s video codec Video Capture Google s video codec 32

33 Workload Analysis Chrome Google s web browser TensorFlow Google s machine learning framework Video Playback Google s video codec Video Capture Google s video codec 33

34 34 TensorFlow Mobile Inference Prediction 57.3% of the inference energy is spent on data movement 54.4% of the data movement energy comes from packing/unpacking and quantization

35 36 Packing Matrix Packing Packed Matrix Reorders elements of matrices to minimize cache misses during matrix multiplication Up to 40% of the inference energy and 31% of inference execution time Packing s data movement accounts for up to 35.3% of the inference energy A simple data reorganization process that requires simple arithmetic

36 36 Quantization floating point Quantization integer Converts 32-bit floating point to 8-bit integers to improve inference execution time and energy consumption Up to 16.8% of the inference energy and 16.1% of inference execution time Majority of quantization energy comes from data movement A simple data conversion operation that requires shift, addition, and multiplication operations

37 37 Quantization floating point Quantization integer Converts 32-bit floating point to 8-bit integers to improve inference execution time and energy consumption Based on our analysis, we conclude that: Both functions are good candidates for PIM execution It is feasible to implement them in PIM logic Up to 16.8% of the inference energy and 16.1% of inference execution time Majority of quantization energy comes from data movement A simple data conversion operation that requires shift, addition, and multiplication operations

38 38 Video Playback and Capture Compressed video VP9 Decoder Display Captured video VP9 Encoder Compressed video Majority of energy is spent on data movement Majority of data movement comes from simple functions in decoding and encoding pipelines

39 39 Outline Introduction Background Analysis Methodology Workload Analysis Evaluation Conclusion

40 We study each target in isolation and emulate each separately and run them in our simulator 40 Evaluation Methodology System Configuration (gem5 Simulator) SoC: 4 OoO cores, 8-wide issue, 64 kb L1cache, 2MB L2 cache PIM Core: 1 core per vault, 1-wide issue, 4-wide SIMD, 32kB L1 cache 3D-Stacked Memory: 2GB cube, 16 vaults per cube Internal Bandwidth: 256GB/S Off-Chip Channel Bandwidth: 32 GB/s Baseline Memory: LPDDR3, 2GB, FR-FCFS scheduler

41 Normalized Energy Normalized Energy Texture Tiling CPU-Only PIM-Core PIM-Acc Color Blitting Compression Decompression Packing Quantization Sub-Pixel Interpolation Deblocking Filter Motion Estimation 77.7% Chrome and 82.6% Browser of energy TensorFlow reduction for Video texture Playback tiling and Mobile Capture and packing comes from eliminating data movement PIM core and PIM accelerator reduces energy consumption on average by 49.1% and 55.4% 41

42 42 Normalized Runtime 1.0 CPU-Only PIM-Core PIM-Acc Normalized Run%me Texture Tiling Color Bli6ng Compression Chrome Browser Decompression Sub-Pixel InterpolaRon Deblocking Filter Video Playback and Capture MoRon EsRmaRon TensorFlow TensorFlow Mobile Offloading these kernels to PIM core and PIM accelerator improves performance on average by 44.6% and 54.2%

43 43 Conclusion Energy consumption is a major challenge in consumer devices We conduct an in-depth analysis of popular Google consumer workloads 62.7% of the total system energy is spent on data movement Most of the data movement often come from simple functions that consists of simple operations We use PIM to reduce data movement cost We design lightweight logic to implement simple operations in DRAM PIM Core Reduces total energy by 55.4% on average PIM Accelerator PIM PIM Accelerator Accelerator Reduces execution time by 54.2% on average

44 Google Workloads for Consumer Devices: Mitigating Data Movement Bottlenecks Amirali Boroumand Saugata Ghose, Youngsok Kim, Rachata Ausavarungnirun, Eric Shiu, Rahul Thakur, Daehyun Kim, Aki Kuusela, Allan Knies, Parthasarathy Ranganathan, Onur Mutlu

45 Backup

46 Video Playback Compressed video VP9 Decoder Display 63.5% of the system energy is spent on data movement 80.4% of the data movement energy comes from sub-pixel interpolation and deblocking filter Sub-pixel interpolation: interpolates the value of pixels at non-integer location Deblocking filter: a simple lowpass filter that attempts to remove discontinuity in pixels 39

47 Video Playback Compressed video VP9 Decoder Display 63.5% of the system energy is spent on data movement Both Functions can benefit from PIM execution They can be implemented in PIM logic Based on our analysis, we conclude that: 80.4% of the data movement energy comes from sub-pixel interpolation and deblocking filter Sub-pixel interpolation: interpolates the value of pixels at non-integer location Deblocking filter: a simple lowpass filter that attempts to remove discontinuity in pixels 40

48 Video Capture Captured video VP9 Encoder Compressed 59.1% of the system energy is spent on data movement Majority of the data movement energy comes from motion estimation which accounts for 21.3% of total system energy Motion estimation: compresses the frames using temporal redundancy between them 41

49 Video Capture Captured video VP9 Encoder Compressed 59.1% of the system energy is spent on Motion Estimation is a good candidate for PIM data movement execution and is feasible to be implemented at PIM logic Majority of the data movement energy comes from motion estimation which accounts for 21.3% of total system energy Motion estimation: compresses the frames using temporal redundancy between them 42

50 TensorFlow Mobile Inference Prediction 57.3% of the inference energy is spent on data movement 54.4% of the data movement energy comes from packing/unpacking and quantization Packing: A simple Quantization: A simple data data re-organization process conversion operation that that reorders elements of matrices converts 32-bit floating point to 8- bit integers 35

51 TensorFlow Mobile Inference Prediction Based on our analysis, we conclude that: 57.3% of the inference energy is spent on Both Functions are good candidate for PIM execution data movement It is feasible to implement them in PIM logic 54.4% of the data movement energy comes from packing/unpacking and quantization Packing: A simple Quantization: A simple data data re-organization process conversion operation that converts that reorders elements of matrices 32-bit floating point to 8-bit integers 32

52 28 Data Movement Study To study data movement during tab switching, we did an experiment: A user opens 50 tabs (most-accessed websites) Scrolls through each for a few second Switches to the next tab We make two key observations: Compression and decompression contribute to 18.1% of the total system energy 19.6 GB of data swapped in and out of ZRAM

53 37 Other Details and Results in the paper Detailed discussion of the other workloads Tensorflow Mobile Video Playback Video Capture Detailed discussion of VP9 hardware decoder and encoder System Integration Software interface Coherence

54 27 Color Blitting Analysis Color blitting generates a large amount of data movement Accounts for 19.1% of the total system energy used during page scrolling 63.9% of the energy consumed by color blitting is due to data movement Based on our analysis, we conclude that: Majority of data movement comes from: Streaming Color blitting access is pattern a good candidate for PIM execution The It is large feasible sizes to of implement the bitmaps color (e.g., blitting 1024x1024 in PIM pixels) logic Color blitting require low-cost computation operations Memset, simple arithmetic for alpha blending, shift operations

55 Data Movement During Tab-Switching Data Swapped Out to ZRAM (MB/s) Data Swapped In from ZRAM (MB/s) GB of data is swapped in from ZRAM and 50 gets 0 decompressed, at a rate of up to 227 MB/s When 0 switching between tabs, compression and decompression Time contribute (seconds) to 18.1% the 250 total system energy GB of data is compressed and swapped 150 out to ZRAM, at a rate of up to 201 MB/s Time (seconds) 32

56 Methodology: Implementing PIM Targets DRAM SoC Vault PIM Core Cache OR... PIM- Accelerator 1 PIM- Accelerator N PIM Logic PIM Logic... PIM Logic Vault Mem Ctrl Vault Mem Ctrl... Vault Mem Ctrl General-purpose Fixed-function PIM PIM core Accelerator A customized A customized low-power in-memory single mm issue logic 2 area core unit budget for the performs a single thread of the PIM logic target A 4-wide SIMD unit 15

57 Cost Analysis: Texture Tiling PIM core SIMD 4-wide Texture Tiling Texture Tiling (int* src, int* dst, ) { for (y = k*columnwidth, y<maxwidth, y++) { int swizzle = swizzle1; mem_copy (dst + (y*swizzle), src + x, k for (x=x1; x<x2; x+=tile) { mem_copy ((x+y)^swizzle, src + x, x3- swizzle ^= swizzle_bit; PIM } accelelerator Requires simple primitives: memcopy, bitwise 0.33 mm operations, mm and simple arithmetic operations 2 (9.4% of the area available per vault ) Multiple In-memory tiling unit (7.1% of the area available per vault ) 57

58 30 Memory Consumption Potential Solution: kill inactive background tabs When the user accesses those tabs again, reload the tab pages from the disk Downsides: 1) high latency 2) loss of data Chrome uses compression to reduce each tab s memory footprint Uses compression to compress inactive pages and places them into a DRAM-based memory pool (ZRAM) When the user switches to a previously-inactive tab, Chrome loads the data from ZRAM and decompresses it

59 Methodology: Identifying PIM 59 targets We pick a function as PIM target candidate if: Consumes the most energy out of the all functions Significant data movement cost (MPKI > 10) Bounded by data movement, not computation We drop any candidate if: Incurs any performance loss when runs on PIM logic Requires more area than is available in the logic layer of 3Dstacked memory